Beruflich Dokumente
Kultur Dokumente
touch@cdr.stanford.edu
Abstract
This paper for the ICRA 2000 Symposium on Dexterous Manipulation presents an overview of research
in dexterous manipulation. We rst dene robotic
dexterous manipulation in comparison to traditional
robotics and human manipulation. Next, kinematics,
contact types and forces are used to formulate the dexterous manipulation problem. Dexterous motion planning is described, which includes grasp planning and
quality measures. We look at various mid- and lowlevel control frameworks, and then compare manipulation versus exploration. Finally, we list what we see as
the current limiting factors in dexterous manipulation,
and review the state of the art and future of the eld.
1 Introduction
The rst step in moving an object from one conguration to another using robotic ngers is to formulate the dexterous manipulation (DM) problem (Figure 1). This problem sets the framework for determining the required actuator forces/torques to produce
the desired motions of the object. In keeping with an
object-centered approach, we work \backwards" from
the object to the manipulators. The development of
the kinematic portion of the DM problem, done here
from force/torque relationships, can also be accomplished from linear and angular velocity relationships.
This development requires knowing the geometric
relationships of the dexterous manipulator-object
system (i.e., contact locations), object geometry,
ngertip and link geometry, and the kinematics of the
1
Joint
object
coordinate
frame
1
2
Contact
T
Force/Torque
Space
space
Jh
Velocity
Space
space
Jh
fc
space
Object
f
m
B
A
contact
coordinate
frame
2.1.1
Jacobian Relationships
= Ji T ftip
(4)
In the DM problem, these individual Jacobians are
brought together to form the hand Jacobian.
= Gftip
(1)
The grasp Jacobian (or grasp map), G, can be obtained by resolving each ngertip force to a common
coordinate frame embedded in the object. For each
ngertip i, this force resolution results in the mapping
matrix Gi .
i
f obj
3
2
1
J1T
6 2 7
6 0
6
7
6
4 ::: 5 = 4 :::
m
0
= Gi ftip
(2)
In Equation 2, the force vectors f are generalized
vectors: they may include both forces and torques.
The individual mapping matrices Gi are concatenated
to form the grasp map G, and the ngertip force vectors are also grouped into one vector.
i
2
fobj
G1
G2
:::
Gm
6
6
4
ftip1
ftip2
:::
ftipm
:::
:::
:::
J2T
:::
:::
32
54
6
::: 7
76
T
Jm
ftip1
ftip2
3
7
7
::: 5
ftipm
(5)
3
7
7
5
minimizing the dimension of the problem. For a detailed treatment of this topic see [19, 28, 30].
The grasp Jacobian developed above allows us to
calculate the required contact forces from the desired
force on the object. In order to produce these forces at
the ngertips, we now develop a hand Jacobian, which
will allow us to calculate the joint torques from the
contact forces [19]. The hand Jacobian, Jh , is based
on the standard Jacobian, which relates end eector
forces to individual joint torques for a robotic manipulator (in this case one for each nger).
Kinematics
fobji
2.1
xc
space
(3)
fobj
=
=
Jh T fc
Gfc
(6)
(7)
Relative velocities
of contact points on
each of the two objects
Jh _ = x_ c
= GT vobj
(8)
The kinematic and force relationships described by
the above equations and Figure 2 are not necessarily
one to one. The system may be over-constrained (i.e.
the ngers may not be able to accommodate or resist all object motions or forces) or the system may
be under-constrained (i.e. there are multiple choices
for nger joint velocities or torques). Typically, an
under-constrained system is desired for dexterous manipulation tasks. The detection of these conditions can
be accomplished by treating the combination of ngers
and object as a parallel-chain mechanism and evaluating the manipulability of the object with respect to the
palm [19]. A summary of kinematic measures useful
in dexterous manipulation is provided in [6].
2.1.2
vx
vy
x
y
z
Velocities of contact
frames on objects
(1 and 2) and spin
Contact
Equations
}u
}u
vx
vz = 0
= 0, vy = 0
!z = 0
!x = 0
Frictionless
point contact
closure grasp and a \manipulable" grasp. Force closure requires only that the ngers can resist an externally applied force, i.e. the opposing resistive force
can be passive or structural. A manipulable grasp requires that the manipulator can actively accommodate
all object motion directions while maintaining contact.
Fn
Ftx
Point contact
with friction
2.3
Fn
If frictional forces are relied upon to achieve a stable grasp, it is important to provide sucient contact
normal forces. To increase these normal forces, an internal force is supplied. This is a vector of contact
force magnitudes that impart no resultant force to the
object and thus lie in the null space of the grasp map,
G. If a vector of contact force magnitudes is decomposed into those producing external forces and those
producing internal forces, each separate force magnitude vector must satisfy the unidirectional constraints
required by the contact types (i.e. ngers can only
push, not pull, on the object surface). In general,
there are many solutions of grasp forces that satisfy
grasp stability while keeping each contact force inside
its friction limits and supplying some internal force.
This leads to an optimization problem, as discussed
in the next section. One treatment of this problem is
the \virtual linkage" concept developed in [41], where
virtual links are imagined between each unique pair of
contact points. The internal forces are then the forces
that each virtual link would experience during the manipulation. For other, more detailed treatments of the
internal/external force concept see [42, 28].
Fty
Ftx
Soft-finger
contact
Mz
Fty
Fn
Fc
Slip
Friction
Cones
Fc
Given a desired object motion, one must rst determine whether the ngers can move the object without
regrasping. This can be accomplished by a reachability analysis [19]. Using the range of possible manipulations for a given hand and object, a workspace can
be constructed that is a function not only of the geometry of the hand and object, but also of the rolling or
sliding that may occur at the contacts. If the ngers
must disconnect with the object in order to move the
object to the desired position, then one can consider
regrasping. A sequence of nger motions and regrasps
are known as a grasp gait [24].
When moving or reorienting an object with a hand,
only a nite amount of local motion can be imparted
to the object before a new grasp must be found. This
can be due to the limited workspace of the ngers of
any hand (human or robot), or collisions among the
nger links, the environment and the object being
manipulated. One method to determine whether local
motions will suce to reorient the object is the grasp
map, a graphical representation of all stable grasps
[24]. In planning, it is also important to realize that
a new grasp cannot always be found if the object
is moved locally until a nger reaches a workspace
limit; often a grasp gait must occur before the limit is
reached.
3.2
Grasp Optimizations
Measures
and
Control Frameworks
The control of dexterous manipulation can be decomposed into three main levels, as shown in Table
2. High-level control includes task and motion planning and grasp choice. Most of these issues have been
discussed already in this paper. Mid-level control includes manipulation phases, for example whether the
ngers are operating independently or cooperatively,
and whether force or motion control is required. Transitions between these phases must be accomplished
smoothly, and events must be sensed to trigger the
transitions. Low-level control includes basic strategies,
(e.g. force or impedance control) and the formulations
of the control problem in the appropriate space (e.g.
the operational space of the grasped object).
Quality
Manipulators used for dexterous manipulation typically have kinematic redundancy. In addition, there
are usually multiple choices for contact locations that
achieve force closure on an object. Therefore, there
can be an innite number of possible grasps for a manipulation. We want to pick the \best" grasp. That
is, we want to choose the optimal contact locations,
contact forces, and nger poses for a particular manipulator, object and task combination.
In order to choose the best grasp, we need to develop a metric that will measure the \quality" of a
given grasp. It is common for this measure to depend on the task requirements. An example of such a
measure was developed by Li and Sastry [25], who separated the task requirements into two parts: wrench
(or force) requirement and twist (or motion) requirement. Each is represented by task ellipsoids, whose
axes indicate the relative magnitude requirements for
the elements of the wrench or twist vector.
While researchers have formulated good conceptual
quality measures for grasps, using these measures for
automatic grasp choice remains dicult. Many successful optimization techniques have been developed
for specifying contact forces given known contact locations and task requirements. Some of these are ecient enough for real time computation (for example,
[4]). Searching for the optimal contact locations is inherently more dicult, because the quality measure is
4.1
Mid-Level Control
The middle level of control for dexterous manipulation has received relatively little attention compared
to the high and low levels. One approach is proposed
by Hyde and Cutkosky [13]. The middle level of control involves the management of the various phases of
a task, as well as the specication of control laws to
be used when transitioning between phases. Research
in neurophysiology has shown that humans grasping
and manipulating objects use a similar approach, receiving signals from specialized skin cells that trigger
shifts between phases of a manipulation task [17].
Dierent phases are likely to use dierent control
laws, as dexterous manipulation is characterized by
changing kinematic and dynamic congurations and
5
Event A crosses
alert confidence
Contact Phase
Event A crosses
commitment confidence
Event B crosses
alert confidence
A
B
Commanded
Velocity
Event
Confidence
Start-up
Action
Interaction
Force
Alert Action
xd
kv (x_
kp
(xg
kv
_ )
x)
xd
(10)
(11)
Low-Level Control
=
=
As shown by Klatzky and Lederman [21], manipulation and exploration go hand in hand. We can obtain
a precise denition for each separately: \Pure" manipulation occurs when the object is completely known.
\Pure" exploration happens when the object is xtured and is not known. Most dexterous manipulation is a combination of manipulation and exploration.
(9)
6
Commands
Control Laws
Desired
Body
Position
Object
Impedance
Controller
Desired
Contact
Locations
Rolling
Controller
Desired
Internal
Force
Internal
Force
Controller
Backward
Grasp
Transform
Forward
Grasp
Transform
Finger
Control
Plant
Hardware
6.2
Software
grasp choice and optimization systems that use multiple ngers are quite slow and the calculations must
be done o-line, particularly when contact locations
must be determined. Another area for improvement
is motion planning. Similar to grasp choice, the algorithms are slow and cannot be accomplished during
the manipulation task. One nal area is the use of
tactile sensing in control. Understanding and using
tactile sensor output for direct servoing has been the
subject of some recent work [33, 27, 43], however, improvements in tactile sensing and data interpretation
are needed to accomplish this in less controlled conditions.
7 Discussion
At present, autonomous, real time dexterous manipulation in unknown environments still eludes us.
In much of the current research, it appears that we
have given up on anthropomorphic hands because of
diculties in hardware development and autonomous
control. The recent trend has been to break the dexterous manipulation problem into small parts that can
be studied separately with specialized hardware. In
many cases the research is done in simulation rather
than experimentally.
Although autonomous dexterous manipulation remains impractical outside of the laboratory,a promising interim solution is supervised manipulation. In
this approach, a human provides the high-level grasp
and manipulation planning, while the robot performs
ne (dexterous) manipulations [32]. Another method
is teaching by demonstration (gesture-based programming) [40]. The human may also perform the interpretation of tactile information in supervised remote
exploration [32].
The miniaturization of manipulation is another area
of with promise. However, manipulations occurring on
a very small scale are dominated by friction and Van
der Waals forces. Stable grasping is often not necessary; the objects will stick directly to the manipulator.
[7] Mark R. Cutkosky and Paul Wright, Friction, stability and the design of robot ngers, International
Journal of Robotics Research 5 (1986), no. 4, 20{
37.
[8] P. Datseris and W. Palm, Principles on the dvel-
opment of mechanical hands which can manipulate objects by means of active control, Journal of
tion of change detection to dynamic contact sensing, International Journal of Robotics Research
13
References
[27]
and force control of robot manipulators: the operational space formulation, IEEE Journal of
[21]
[22]
[23]
[24]
[25]
[26]
trol of 3-d rolling contacts in two-arm manipulation, IEEE Transactions on Robotics and Au-
10